1 /* { dg-do compile } */ 2 /* { dg-options "-O2 -ftree-vectorize -msve-vector-bits=scalable" } */ 3 4 #include <stdint.h> 5 6 #define VEC_PERM(TYPE) \ 7 TYPE __attribute__ ((noinline, noclone)) \ 8 vec_slp_##TYPE (TYPE *restrict a, int n) \ 9 { \ 10 for (int i = 0; i < n; ++i) \ 11 { \ 12 a[i * 2] += 10; \ 13 a[i * 2 + 1] += 17; \ 14 } \ 15 } 16 17 #define TEST_ALL(T) \ 18 T (int8_t) \ 19 T (uint8_t) \ 20 T (int16_t) \ 21 T (uint16_t) \ 22 T (int32_t) \ 23 T (uint32_t) \ 24 T (int64_t) \ 25 T (uint64_t) \ 26 T (_Float16) \ 27 T (float) \ 28 T (double) 29 30 TEST_ALL (VEC_PERM) 31 32 /* { dg-final { scan-assembler-times {\tmov\tz[0-9]+\.h, w[0-9]+\n} 2 } } */ 33 /* { dg-final { scan-assembler-times {\tld1rw\tz[0-9]+\.s, } 3 } } */ 34 /* { dg-final { scan-assembler-times {\tld1rd\tz[0-9]+\.d, } 3 } } */ 35 /* { dg-final { scan-assembler-times {\tld1rqd\tz[0-9]+\.d, } 3 } } */ 36 /* { dg-final { scan-assembler-not {\tzip1\t} } } */ 37 /* { dg-final { scan-assembler-not {\tzip2\t} } } */ 38 39 /* The loop should be fully-masked. */ 40 /* { dg-final { scan-assembler-times {\tld1b\t} 2 } } */ 41 /* { dg-final { scan-assembler-times {\tst1b\t} 2 } } */ 42 /* { dg-final { scan-assembler-times {\tld1h\t} 3 } } */ 43 /* { dg-final { scan-assembler-times {\tst1h\t} 3 } } */ 44 /* { dg-final { scan-assembler-times {\tld1w\t} 3 } } */ 45 /* { dg-final { scan-assembler-times {\tst1w\t} 3 } } */ 46 /* { dg-final { scan-assembler-times {\tld1d\t} 3 } } */ 47 /* { dg-final { scan-assembler-times {\tst1d\t} 3 } } */ 48 /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.b} 4 } } */ 49 /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.h} 6 } } */ 50 /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.s} 6 } } */ 51 /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.d} 6 } } */ 52 /* { dg-final { scan-assembler-not {\tldr} } } */ 53 /* { dg-final { scan-assembler-not {\tstr} } } */ 54 55 /* { dg-final { scan-assembler-not {\tuqdec} } } */ 56